A FENN-TL Approach for Reliability Analysis of a Primary Ice Detection System

Abstract

Solving the reliability problem of primary ice detection systems is of great significance to support the design of anti-icing systems. In this paper, an efficient method employing a feature-enhanced neural network (FENN)–transfer learning (TL) surrogate model was developed to process two types of features (flight and aircraft parameters). A FENN was established with an autoencoder, and TL was implemented with 15 new points. A new loss function was designed and combined with FENN to control the direction of prediction error. The determination coefficient was 0.993 in the holding state and 0.997 in the local area near the dangerous state. Based on 1 million predicted results of Common Research Model (CRM) airfoil, the primary ice detection system is most likely to have reliability problems at a low angle of attack and low-speed flight state, and angle of attack has the greatest influence. FENN-TL proved a flexible and efficient method for reliability analysis of primary ice detection systems. This method and the obtained CRM results can be further used to support the design and airworthiness certification of large aircraft.